GenAI Use Safety Modes

Rank GenAI collaboration from checking your work up through independent complex problem-solving — autonomy rises with risk and oracle demand.

Bach & Bolton rank GenAI use by safety:

Check your work — ask for critique of an answer you already hold (safest).
Assist solving a problem — draft a strategy or draft code while you stay alert.
Independently perform a specific procedure with verifiable behaviours — needs agentic capability; significant risk.
Independently solve a complex problem — highest risk; weakest oracles.

Map this onto progressive agent architecture and human approval gates: climb the ladder only when you have offline oracles and blast-radius limits for the new autonomy. Wishful prompt engineering without a stable evaluation contract leaves modes 3–4 untestable — pair with genai radical fragility and offline prompt evaluation.

Sources